Low-power Precision Assignment for Special Function Instructions of 3-D Graphics Processing Units

• Main Authors: Wei-Chung Chi, 姬瑋忠
• Other Authors: Shiann-Rong Kuang
• Format: Others
• Language: zh-TW
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/74872065376416598206

碩士 === 國立中山大學 === 資訊工程學系研究所 === 101 === Today COMS process and 3-D graphics technology are quickly developed, and 3-D graphics processing units (3-D GPU) have already been used in the handheld devices. However, this application requires many complex operations which consume a lot of power. This is a severe challenge for handheld devices with limited battery capacity. Furthermore, the development of battery manufacturing technology is slower than the IC process technology. As a result, the power storage of handheld devices cannot supply for a long time of operation. If we want to effectively use the limited battery capacity, the power-saving mode must be added to the application. In this thesis, we propose a low-power precision assignment for special function instructions in 3-D GPU. Its special function operations include reciprocal, reciprocal square root, logarithm and exponential. Each operation is able to perform one of four different precision modes dynamically in demand. In this way, not only some power consumptions can be saved, but also the quality of graphics is able to be accepted by users. This paper first introduces a multi-precision function interpolator which is compliant with the IEEE-754 single precision floating point standard, and this function interpolator is used for special function operations in 3-D GPU. This multi-precision function interpolator provides four precision modes to trade power consumption with output accuracy of special function operations. Subsequently, affine arithmetic is modified to build the error model of special function. This error model indicates the relationship between the precision mode of each special function operation and the final output. Based on this multi-precision function interpolator and the modified affine arithmetic error model, a fast tabu search (TS) algorithm is developed to assign the precision modes of each special function operation. Besides, the necessary graphics output quality can be provided, according to the requirements of the users.